ADVANCED SATELLITE-BASED OIL SPILL DETECTIONSPILL DETECTION

INSTITUTO HIDROGRÁFICO, LISBON, 24-26 JUNE 2014

Andrea Radius*1, Rui Azevedo*1, Delfim Rego*1 , Paulo Carmo*1

*1 Edisoft, Paço de Arcos, Portugal

Maritime awareness

• Ocean health is crucial for the Earth well being

• The ocean is a eco-system with major responsibility

for climate, coastal erosion, etc

• Maritime pollution monitoring became an essential

tool to monitor sea health and a very important

challenge in terms of natural environmental

preservation

Maritime Monitoring context

• Sea pollution monitoring importance is stressed by

several European (and not only) projects that aim on

sea preservation

• The synergy between R&D and operational

projects is crucial to implement global monitoring projects is crucial to implement global monitoring

architectures in the near future

• EDISOFT works on the bigger European operational

project of pollution monitoring: EMSA CleanSeaNet

• Other important R&D projects which EDISOFT have

worked: EC FP7 Sea-U and ESA AOSD

•

• The knowledge acquired during the aforementioned

projects allowed EDISOFT to plan the implementation of

a non-supervised automatic tool for oil spill detection

AOSD context

a non-supervised automatic tool for oil spill detection

• The project involves strong R&D components to

improve oil spill detection capabilities

• Convergence to an automatic algorithm

• Being closely related to the operational segment

• Focused on the oil spill detection processing chain

automation (e.g. CleanSeaNet)

• Integration with an already tested and optimized

AOSD objectives

• Integration with an already tested and optimized

software as NEST

• Integrating auxiliary data (like wind data), together

with image and geometry analysis

Software architecture

Level 1BLand

Mask

Data

Reader

Filter

WindsWind

The AOSD follows the processing model of NEST; its

architecture is the following.

Statistical

Analysis

Dark

Pixels

Detection

Vectorize

Clusters

Cluster

Dark

Pixels

Wind

Trainset

Contours

Filtering

Features

Extraction

ClassifyOil Spill

Report

Data reader

The current software version reads data acquired

from the following satellites:

• Envisat

• Radarsat-1

• Radarsat-2• Radarsat-2

• Imagery data can be combined, if available, with

winds ancillary data

Land mask

• Land areas are masked with no data value, to be

excluded by the detection component

• A user defined buffer is added to exclude areas

near the coast

Filter winds

• Invalid wind areas are excluded from the analysis

• “Invalid” is defined according to the user input

criteria

� Low winds – lower � Low winds – lower

values are excluded

� High winds – higher

values are excluded

Statistic analysis

• Smaller window areas are analyzed to find

individual threshold values

• Threshold values are automatically adapted to a

sliding-window covering each analysed area

� Window size – defined by � Window size – defined by

the user

� Statistic data extraction – in

each sliding-window

� Threshold – combining

statistical data

Dark pixels detection

• Each pixel is filtered according to the established

window threshold

• Only lower amplitude values allow the analysed

pixel to move on to the next phase

Cluster dark pixels

• Darker pixels are

aggregated according to

a user defined distance

• Each group of

aggregated pixels form aggregated pixels form

a cluster

• Non-clustered pixels

are discarded

Vectorize clusters

• Clusters are

vectorised to

polygons

• The remaining dark clusters are converted from

raster to the vectorial domain

• Polygons are

simplified to

contours

polygons

Contours filtering

• Formed contours can be discarded if they do not

comply with three exclusion criteria:

1. Being an invalid polygon1. Being an invalid polygon

2. Dimension

3. Closeness to an invalid wind area

Features extraction

• Well defined features that can better characterise

an oil spill were previously chosen

• Based on the definition, the individual features of

each contour are extracted

• The individual feature values are delivered to the

classifier

Classify

• All detected and filtered remaining patterns are

then classified based on their individual features

• Classification is done in two levels

Contrast

1st Level 2nd Level

Features

Spill

Look-alike

Contrast

Type

Shape

Surround

Edges

Classify

• The first classification level is based on the Support

Vector Machines (SVM).

It has the purpose to classify the detected dark

patch as oil spill or look-alike.

• The second level of classification is executed for

the classification of the detected oil spills, with the

purpose to characterize the oil spill according to its

attributes

Oil Spill Report

• Available in three different formats:

1. Esri shapefile

2. EMSA OSN

3. Google Earth KMZ

Correlation analysis

• Classification can only provide efficient predictions if the model is well balanced

• Development of a prediction model capable of classifying dark features as a spill or a look-alike

• It was essential to determine which variables are relevant to the model

• Understanding each variable relevance was performed through a correlation analysis

Oil spill characteristics

• Type – linear, tail, angular, patch, droplet

• Contrast – strong, medium, weak

• Surround – homogeneous, inhomogeneous

• Shape – smooth, irregular

• Edges – Sharp, sharp and diffuse, diffuse

• Wind speed – low, moderate, high

Extracted variables

• The variables are extracted according to the

aforementioned oil spill characteristics

• Definition of each detection with the variables:

� Area

� Perimeter

� Length

� Width

� Linearity

� Complexity

� Density

� Gradient

Decision variable

• Spill confidence level

• Divided in five sub-level types:

� Ongoing spill

� High confidence

� Medium confidence

� Low confidence

� Lookalike

• The variable holding the prediction result

Results

• Prepared dataset based on the available archive

• Total of 78 ENVISAT scenes randomly selected

• AOSD validation data set results:

AOSD Lookalike Low Medium High Total

Success 18585 76 71 41 18773

Fails 3003 56 58 30 3147

Rating 86% 58% 55% 58%

Overall performance 86%

Detection example I

Detection example II

CSN results (2007-2011)

Future work

• Upgrade to the latest available version of NEST

• Integration of AOSD within the EMSA CSN service

• Integration of other SAR systems – Sentinel-1,

TerraSAR-X, COSMO-SkyMed

• Spill orientation – to detect possible responsible ships• Spill orientation – to detect possible responsible ships

• Integrate other ancillary data sources – to help

preventing false positive detections

• Improve time efficiency – to be used operationally

• Incorporate additional layers of information – as

detected vessels and AIS positions

THANK YOU

Rua Calvet de Magalhães, N245

2770-153 Paço de Arcos

Portugal

TEL: +351 212 945 900

FAX: +351 212 945 999

www.edisoft.pt

edisoft@edisoft.pt